Accurate where the "0.81% error in 12 KB" quote says it is — 0.754% RMSE at one million visitors over 30 trials. But the same sketch was 2.6% off at 40,960 visitors, where it switches estimators, and 89% low when the hash was FNV-1a on sequential user IDs.
Recursive shadowcasting computed field of view in 4.5 µs at radius 16 on a 256x256 map with 10% walls, about 3x faster than a Bresenham line to every cell, and missed 0 of 6,759 visible cells where the lines missed 959. The version most people copy has a bug that shows 7x more cells through solid wa
We fed 48 timestamp shapes from real APIs to new Date() under three server timezones. It returned the wrong instant for 4, 6 or 9 of them depending only on TZ, and Invalid Date for 11 more. date-fns and luxon disagree with it on date-only strings.
No. Tested on 88 hand-tagged posts from this blog, TF-IDF's top five terms recovered 10.0% of the human topic tags at best, and the textbook log-scaled variant 5.1%. 27.6% of the tags never appear in the post at all. The same vectors picked the category correctly 84.1% of the time.
Across 1,000 seeded 200x120 maps, cellular automata caves came out fully connected 0.5% of the time. BSP, random room placement and drunkard's walk were connected every time. But the caves' stray pockets held about 1% of the floor and took 16 carved cells to fix, while discarding and retrying failed
Yes, by about 3x on HTTPS. At 380 ms RTT, a reused connection answered in 385 ms against 1,153 ms for a new one, measured with server-side connection counts. The catch is a stale-socket race that cost 14 to 19 errors per 1,000 requests until we fixed the idle timeout.
Trigram Jaccard on normalised names had the best F1 in our test, 0.914 against 0.767 for Levenshtein, over 31,342 labelled pairs of Vietnamese and English names. But at that threshold it merged 100% of 5,457 pairs of different Vietnamese people who share a family and middle name. On a synthetic 10,0
The best of four commonly copied phone regexes found 28.94% of 3,801 labelled phone numbers, at F1 0.4356. libphonenumber-js reached F1 0.8534. A regex tuned for Vietnamese and international formats beat it on the formats it was written for, then found 22.79% of the formats it was not.
A 128x128 chunk of 4-octave noise took 0.30 ms to 1.92 ms in Node on an Apple M3, depending on the noise. Simplex never beat branch-free Perlin, a switch statement made 8 octaves 4.7x slower, and calling one chunk loop with two noise functions made it 3x slower.
Yes. Over 10,000,000 rolls Math.random hit a 1% drop 0.9998% of the time, as close as crypto.getRandomValues. The streaks players call broken are the maths of independent rolls — 13.5% of 100,000 simulated players got nothing in 200 kills at 1%, matching the 13.40% formula. A pity timer fixes that;
About 97% of it. One second of a 30-bone walk cycle at 60 fps is 72,005 bytes of raw float32 and 2,229 bytes after constant-channel removal, cubic keyframe reduction and quantisation — 32x, with 7.2 mm of worst-case joint error on a 1.8 m character. Measured on a walk cycle, a snappy jump and a jitt
A newly deployed subdomain took 8.1 seconds to first byte, then was fast. We broke container cold start into five measured phases. The whole in-process budget is 448 ms. The seconds live outside the process — in the image pull and the healthcheck interval.
Infrastructure
No. On a small flat object structuredClone runs at 1.28M ops/sec against JSON's 2.07M, and a five-line hand-written deep copy beats both at 8.71M — 6.8x faster than the native call. Use structuredClone anyway, because the fast one silently turns a Date into an empty object and stack-overflows on a c
Games
A failed A* search on a 256x256 grid expands 52,286 nodes — 99.7% of every walkable cell on the map — against 10,920 when a path exists. It costs 7.57 ms instead of 1.66 ms. The heuristic that makes A* fast contributes nothing when the goal is unreachable, because there is no goal to be guided towar
Games
At 100,000 sprites, sorting a shuffled list costs 27.0 ms per frame — 162% of a 60fps budget, before you draw anything. Sorting the same list when it is already in order costs 1.35 ms, and 3.36 ms when 1% of sprites have moved. The sort is not the problem. Rebuilding the list every frame is.
AI & Data
Measured in SQLite FTS5 on Apple M3. A full rebuild of a 100,000-document index costs 1.36s; updating one document costs 0.10ms. Incremental wins until about 9% of the corpus changes in a day at 100k documents, and only 5% at 1M. Plus what heavy churn does to index size and query latency, and what a
Games
Not below about 2,000 spawned objects per frame, and the free-list pool everyone writes made the median frame 2.5x slower at 3,000 spawns and 4.8x slower at 10,000. Allocating a bullet costs 3.55 ns more than reusing one; the pool's own push and pop cost 7.86 ns. What pooling can buy is the tail — b
Integration
We applied 12 realistic breaking changes to a live API and measured which detection method caught each. Runtime validation caught 7 of 12, a recorded golden response 9, property checks 9, TypeScript 0. Validating every response costs about 20 microseconds.
Integration
You send it all again. A 256 MiB single POST killed at 90% wasted 242,221,056 bytes and needed 510,656,512 bytes in total to succeed; the chunked version re-sent 419,430 bytes — 577x less. Measured on Node 23: chunk size against failure rate, resume discovery, per-chunk checksums, parallel chunks, t
AI & Data
A PDF page has no paragraphs, no sentences and often no spaces — it has draw-glyph-here instructions. Sorting those by position, the obvious extraction algorithm, put 0 of 23 consecutive sentence pairs next to each other on our two-column test page. macOS PDFKit got 23 of 23, then broke a letter-spa
Games
The naive version — full state as JSON, every tick — costs 1,101,710 bytes per second per client at 128 entities and 60 Hz, and 141 Mbit/s of server upload for 16 players. Binary packing, 16-bit quantisation, delta encoding and interest management together brought that to 70 payload bytes per tick,
Infrastructure
About 2 microseconds of CPU and 238 bytes of disk per line — and five log lines per request cut a Node 23 server from 68,722 to 46,059 requests per second. Measured: cost per call by log style, the argument-evaluation trap that costs 1,859 ns per disabled debug call, whether stdout blocks the event
AI & Data
A ZIP of XML parts — 17 of them in the document measured here. The trap is the paragraph model. The sentence "Send the invoice before Friday." is stored as 5 separate w:r runs because one letter is bold, and a naive strip-the-tags extractor reads "The unit price is 100200 USD" from a document that s
Games
There is no crossover. A struct-of-arrays layout beat an array of plain objects at every entity count measured, including 1,000 — but at 1,000 it saved 0.005 ms per frame. The saving first reaches a whole millisecond at 184,000 entities. And the cache-locality argument everyone repeats reverses on i